feat(agentic-harness): make generated orchestrator the hard-gate entry point (v0.6.0)

[MrBogomips]

Summary

harness-setup generates a per-project orchestrator skill and registers a pointer in the target repo's CLAUDE.md. Until now that pointer carried only a soft, advisory trigger line:

Trigger: For {domain} work, use the {orchestrator-skill-name} skill. Answer simple questions directly.

So activation rested on (a) the model voluntarily heeding a soft sentence and (b) probabilistic skill-description matching — which the harness's own skill-writing-guide.md admits is conservative ("the model tends to skip a skill for a task it thinks it can handle bare-handed"), and maintenance.md even lists "the user bypassing the orchestrator" as a known failure mode. The orchestrator was not reliably the entry point.

This PR makes the generated CLAUDE.md register the orchestrator as a hard gate: the model must invoke it before responding to any prompt, with no CLAUDE.md-level bypass. CLAUDE.md is the highest-priority instruction layer, so a firm standing directive there is the reliable lever — not the probabilistic description. The triage that used to live in CLAUDE.md ("answer simple questions directly") moves inside the orchestrator's first phase (Phase 0: intake & triage), so a trivial or off-domain prompt still routes through the orchestrator and is answered quickly there, without spinning up a team for every message.

This is a behavior change to generated output only — minor and backward-compatible. Existing harnesses keep working; re-running setup/review upgrades them.

User decisions

• Routing model = hard gate on all prompts (no escape valve at the CLAUDE.md layer).
• Enforcement = CLAUDE.md directive + strengthened skill description; no hook (stays within the harness's current footprint).

Changes (9 files)

• shared/claude-md-pointer.md (primary) — replaces the soft trigger with an entry-point directive (hard gate); relocates the triage note into the orchestrator; adds a multi-harness composition note (write the gate once as a routing preamble, each section keeps an Orchestrator: line).
• skills/harness-setup/references/orchestrator-template.md — orchestrator descriptions in templates A/B/C now open with "Entry point for all {domain} work…"; a Phase 0: intake & triage front-step added to all three; SDD-coordination addendum + worked snippets re-headed to match; follow-up-keywords note.
• skills/harness-setup/SKILL.md — Step 5 ("register the pointer") wording, the build-into-orchestrator bullets, and two new deliverable-checklist items (description asserts entry-point status; first phase is intake & triage).
• Consistency — harness-review/SKILL.md, harness-review/references/usage-assessment.md, harness-setup/references/maintenance.md, shared/harness-model.md: "trigger rule" → "entry-point directive"; harness-review now flags a soft gate (or a bypassed orchestrator) as a finding.
• Version/docs — plugin.json 0.5.0 → 0.6.0; README line.

Test plan

• bash tests/ci/run-structural-tests.sh — all suites passed (6 pre-existing warnings, none in agentic-harness)
• claude plugin validate ./agentic-harness --strict — passed
• Stale-phrasing sweep (grep -rn -i "answer simple questions directly\|trigger rule" agentic-harness/) — only the one intentional explanatory note remains
• Template read-through: pointer + Phase 0 render coherently; SDD Phase 0 addendum still composes after the triage front-step

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.